SmartSweepers version Py | ======================== | Author: Yati Sagade | | Inspiration: Mat Buckland(Fup)'s C++ code | at http://www.ai-junkie.com | | Mail: yati dot sagade at gmail dot com | --------------------------------------------+

This is my take on the classic example of the workings of an Artificial Neural Network, originally found at http://www.ai-junkie.com/ The original(Fup's) code was C++/Win32.

Well, if you don't know already, here's what we're trying to achieve: Using ANNs and learning through the Genetic Algorithm(GA), breed a set of mine- sweepers that wander around gobbling up any mines in the world. The total number of mines in the world at any instant is constant - so when a mine is swept, another appears somewhere in the world. The fittest minesweepers are displayed in a different colour (I prefer green for the fittest, and blue for the rest).

i) Install Python 2.x if you haven't done so already. ii) On the shell, cd to this directory. iii) Make sure you have pygame installed. On the prompt, say

    $ pip install pygame

  Ofcourse, you'll need to have pip installed. For that,
    
    $ yum install pip

  or your system's equivalent pkg-mgmt command should do the job.

iv) type

    $ python main.py
        
        OR

    $ ./main.py

For the graphical rendering, the fantastic pygame library has been used, and hence, of course, pygame is needed to run the program.

Lowdown on the files in this project:

genetic_algorithm.py The Genetic Algorithm module, should be pretty easy to understand provided you've been reading up on Fup's excellent tutorial on the topic.

neuralnet.py Home for the Neuron, NeuronLayer and NeuralNet classes.

minesweeper.py provides the Mine and MineSweeper classes.

geom2D.py This is the mathematical grit that goes behind all the fancy stuff :) It is home to the Matrix2D, Vector2D and PointList classes. PointList subclasses Matrix2D and is used to store the vertices of our sweepers and mines as a matrix of 2D points. We only needed multiplication of matrices for the basic transforms using normalized co-ordinates, so only Matrix2D.mul() is implemented leaving aside add, sub or div.

If you don't know what normalized co-ordinates are, panic not. They're just
a way to represent points such that transformations(like rotation, 
translation, etc) are reduced to a chain of matrix multiplications.
This requires the addition of an extra co-ordinate(yes, that's 3
co-ordinates for 2D points). This extra co-ordinate is almost always chosen
to be fixed at unity(1). So, (0,0) becomes (0,0,1) and (2.75,5) becomes 
(2.75, 5, 1) and so on.

The benefit is that by defining transformation matrices - one for
rotation, one for translation, one for scaling and so on - that respect the
normalized co-ordinates, we can effectively express every transformation as
a matrix multiplication, thus allowing for chaining transforms. There's even
an algorithm to find the most efficient order of multiplying matrices given 
a matrix multiplication expression by exploiting associativity of matrix 
multiplication.

settings.py Global settings - Equivalent of CParams.cpp and params.ini in Fup's code.

utils.py miscellaneous utility functions

controller.py The "duct tape" that glues together all other components of this project. it takes care of creating the minesweepers, breeding generations of them and rendering them.

main.py This just creates the pygame window, and loops throught the event loop while invoking a Controller instance's functions to update and render the world.

resources/ Currently contains the icon I ripped from Fup's code(that was originally in .ico format, I made a GIF of it - hopefully not offending anyone).

TODO: (i) Any improvement in performance (ii) Make the code conform to PyLint conventions (currently the score is really bad!).